Qwen3-Next-80B-A3B-Thinking

Qwen3-Next-80B-A3B-Thinking is the reasoning variant of the Qwen3-Next series from Alibaba's Qwen team, released 2025-09-10 under Apache 2.0 with open weights. It shares the instruct model's hybrid attention stack (Gated DeltaNet plus Gated Attention for ultra-long context) and a high-sparsity Mixture-of-Experts layout with 80B total and 3B active parameters, and was post-trained with GSPO to stabilise reinforcement learning over that hybrid architecture. It uses roughly 15T training tokens and is served by Novita with a 65K-token context window.